Capability 03 / 04 · The model layer of the brand

A model that already knows what the brand looks like.

Brand AI Tools builds the custom-tuned models behind brand delivery. Image, product and language models trained on your own rights-cleared material with Flux, Adobe Firefly, ComfyUI and open-weight families, scored against a brand fidelity set, and served to the tools your teams already use.

Start a brand model brief What it covers
Typical length
6–10 weeks to a tuned model
Model types
Image · product · language
Ownership
You own the weights

Illustration: a custom-tuned brand model. A Flux base model with a brand adapter, trained on 1,240 approved assets, is given the prompt "autumn window, ceramic range, three variants". It makes three variants. A brand check scores each on colour difference, type and logo clear space: variants A and B pass at 0.93 and 0.90 and are approved; variant C fails colour and clear space at 0.78 and is sent back to regenerate.

What it covers

Six parts of a brand that a model can learn

Tuning is the small part. The work is the dataset, the rights position, the scoring and what happens the day the brand changes.

  1. 01 / 06

    Rights-cleared training sets

    Your assets curated, labelled and checked for licence and consent item by item, with a held-out set kept back for scoring.

    Curated · consented

  2. 02 / 06

    Brand-tuned image & product models

    Adapters and fine-tunes on Flux, Firefly or open-weight families, so generation starts on brand, including packshots and product likeness.

    LoRA · fine-tune

  3. 03 / 06

    Brand language models

    Tuned prompts or a fine-tuned model that holds the voice, the lexicon and the words you never use, inside the tools your team writes in.

    Voice-aware

  4. 04 / 06

    Brand fidelity evaluation

    A fixed scoring set for palette, mark use, typography, composition and tone, run on every model version before it is released.

    Scored every version

  5. 05 / 06

    Guardrails & approval

    What the models may generate, what needs a person, and prompt-injection and misuse testing before the tools reach a team.

    Human in the loop

  6. 06 / 06

    Serving, versioning & care

    A model registry, versioned releases, cost and latency limits, and re-tuning when the brand moves.

    Registry · re-tune

How it runs

Ground, tune, score, then serve

The brand system becomes a dataset and a scoring set. Nothing is released on a subjective look at four good images. Timings are typical, set per engagement.

  1. Wk 01–03

    Ground

    Assets, rules and counter-examples collected, labelled and rights-checked. The brand fidelity scoring set is written with your brand team.

    • Training set
    • Rights register
    • Fidelity scoring set
  2. Wk 03–06

    Tune

    Adapters and fine-tunes trained and compared on the held-out set. Base model, method and settings chosen on the scores.

    • Model candidates
    • Score report
    • Model card
  3. Wk 05–08

    Guard

    Guardrails, prompt-injection and misuse tests, an approval queue and an output log, all in place before a team touches it.

    • Guardrails
    • Red-team report
    • Approval workflow
  4. Wk 08–10

    Serve

    The model deployed behind an API and into your design tools, with a registry, cost limits and a re-tuning plan.

    • Served model
    • Model registry
    • Care plan

In practice · model card

A model ships only when it beats the base model on your brand.

The release gate compares the tuned model with the stock model it started from, on the checks your brand team cares about. Each has a threshold; the card says pass or hold.

Model card · Brand image model v3Release gate
Model card · Brand image model v3: illustrative sample
CheckBase modelTuned v3ThresholdGate
Palette drift (mean ΔE 2000) 5.81.7≤ 2.0Pass
Brand-rule adherence 46%88%≥ 85%Pass
Product shape fidelity 0.710.93≥ 0.90Pass
Unsafe or off-policy outputs 1.9%0.3%≤ 0.5%Pass
Text rendering in image 38%61%≥ 80%Route to designer
Training images with provenance
100%Licensed or owned
Cost per approved asset
Index 45Base model = 100
Checks passed
4 / 5Fifth stays with people

Illustrative Thresholds are agreed with your brand team before training starts.

What changes

On-brand at the first attempt, at a fraction of the cost.

Measured on every generated variant against your brand palette and rules. Typical targets, not promises; yours are set per engagement.

01

On brand before anyone edits it

Generation starts from your identity, and every version is scored against the same fidelity set before it is released.

On-brand at first go

  • Before40%
  • Target85%+

Measured by Brand check on every generated variant

02

A dataset you are allowed to use

Licence and consent recorded per item, counter-examples included, and nothing scraped in hope.

Colour drift (ΔE)

  • Before5.8
  • Target≤ 2.0

Measured by Mean ΔE 2000 against brand palette

03

Yours, and portable

Weights, datasets, prompts, scores and logs sit in your accounts. Nothing is retained and nothing else is trained on them.

Cost per approved asset

  • BeforeIndex 100
  • TargetIndex 45

Measured by Model spend + review time, indexed

What you keep

Your tuned models, weights and all.

Model weights, training sets with provenance, evaluation reports and the serving configuration sit in your accounts. Nothing is held hostage to a platform.

Colour swatch cards fanned into a full circle on a grey surface
The brand palettewhat every output is scored against

Manifest · Brand Models7 items

Brand AI Tools deliverables and their formats
#DeliverableFormat
01Rights-checked training & evaluation setsDataset · register
02Tuned model weights or adaptersWeights · your accounts
03Model card & training documentationDoc
04Brand fidelity scoring set & reportTests · report
05Guardrail & red-team test resultsTests · report
06Serving endpoint & model registryAPI · registry
07Output log & approval workflowDashboard · workflow

Two tools, one name

This one makes. Brand Design’s one checks.

Brand AI Tools is a capability of both disciplines. Here it means custom-tuned generative models that produce on-brand work. In Brand Design it means the governance tooling that checks any asset against the brand system, whoever or whatever made it.

  1. Brand Design Brand system

    Identity, voice, tokens and rules.

  2. AI Design · Brand AI Tools · this page Custom-tuned models

    Trained on approved work, making on-brand variants.

  3. Brand Design · Brand AI Tools Governance checks

    Every asset scored against the system, from any source.

  4. Your team Approval

    A named person signs off. The record is kept.

They meet at the scoring step: our models are tuned until they pass the checks Brand Design wrote.

See Brand Design’s Brand AI Tools

Technologies we work with

Trained on open tooling, served where you choose.

The training, evaluation and serving tools brand models are built with. Weights and configuration stay portable between providers.

  • PyTorch
  • Hugging Face
  • Replicate
  • Modal
  • Python
  • MLflow
  • ONNX
  • NVIDIA
  • vLLM
  • Anthropic
  • Google Gemini
  • Figma
  • GitHub

Also in use

  • OpenAI

Frameworks we build to

Trained on data you may use, governed like any other system.

The frameworks behind the training-data record, the risk register and the release gate. We build to them; they are not certifications we hold.

  • ISO/IEC 42001:2023Artificial intelligence management systems

    Requirements for establishing, running and improving an AI management system: AI policy, impact assessment, data and lifecycle controls.

  • NIST AI RMF 1.0AI Risk Management Framework

    Four functions for trustworthy AI: Govern, Map, Measure and Manage, with a companion profile for generative AI (NIST AI 600-1).

  • EU AI ActArtificial Intelligence Act (EU) 2024/1689

    A risk-based regime: prohibited practices, obligations for high-risk systems, transparency duties and rules for general-purpose AI models.

  • OWASP Top 10 for LLM ApplicationsLLM and generative AI security risks

    Risks specific to LLM systems, including prompt injection (LLM01), sensitive information disclosure (LLM02), excessive agency (LLM06) and vector and embedding weaknesses (LLM08).

  • GDPRGeneral Data Protection Regulation (EU) 2016/679

    Lawful basis, data-subject rights, data protection by design and by default, breach notification and DPIAs for high-risk processing.

  • DPDP Act 2023Digital Personal Data Protection Act, 2023

    Notice and consent, duties of data fiduciaries, rights of data principals, breach intimation and added duties for significant data fiduciaries, with the DPDP Rules.

Services & packages

Tuned on your material. Scored before release. Owned by you.

Buy one model, such as a brand image model or a voice model, or the whole model layer: the dataset, the tuning, the scoring, the guardrails and the serving. Weights, data and logs are delivered into your accounts, and a person approves what ships.

Categories
04
Services
14
Packages
04
Not sure what you need? Describe the problem

How to buy

  1. 01Pick services. Enquire about one, or add several to a brief.
  2. 02Choose how to engage. A sprint, a fixed project or an ongoing team.
  3. 03Send the brief. We reply within one working day.

Browse by category

Timelines are typical. Every quote follows a written scope.

01Data & rights

3 services
Typical timeline: 2–5 weeks

Rights-cleared training set

Your assets turned into training data you are allowed to use: curated, labelled and checked for licence and consent, item by item.

What’s included

  • Asset collection, selection and labelling
  • Licence and consent check per item
  • Counter-examples and a held-out set
  • Dataset versioning and documentation
  • Data
  • Rights-checked
  • Versioned

Best forAny brand preparing to tune a model on its own material.

Typical timeline: 3–5 weeks

Brand knowledge base for AI

Your guidelines, rules and approved examples structured so AI tools can read and cite them, and kept current when the brand changes.

What’s included

  • Guidelines turned into structured rules
  • Searchable index of approved examples
  • Connectors into your tools
  • Update process for brand changes
  • Retrieval
  • Cited
  • Current
  • LlamaIndex
  • pgvector

Best forBrands whose rules sit in PDFs that no tool can read.

Typical timeline: 2–4 weeks

Model & platform selection

Which base model and which tuning method for each brand task, chosen by testing on your own assets rather than on benchmark scores.

What’s included

  • Shortlist of base models per task
  • Side-by-side trial on your own assets
  • Licence, data-residency and cost review
  • Recommendation with a rationale
  • Vendor-neutral
  • Tested
  • Cost
  • OpenAI

Best forTeams weighing base models and platforms before committing.

02Tuning

4 services
Typical timeline: 4–8 weeks

Brand-tuned image model

An image model tuned on your identity with Flux, Firefly or an open-weight family, so generation starts on brand instead of drifting toward it.

What’s included

  • Adapter or fine-tune on your training set
  • Comparison across methods on a held-out set
  • Brand fidelity score report
  • Weights and prompts delivered to your accounts
  • LoRA · fine-tune
  • Scored
  • Yours

Best forBrands generating imagery at a volume no one can check by hand.

Typical timeline: 4–8 weeks

Brand voice & copy model

Tuned prompts or a fine-tuned language model that holds the voice, the lexicon and the words you never use, inside the tools your team writes in.

What’s included

  • Voice and lexicon turned into structured rules
  • Tuned prompts or a fine-tuned model
  • Scoring against the voice rules
  • Integration with your writing tools
  • Voice-aware
  • Copy
  • In your tools
  • OpenAI

Best forTeams writing at volume across markets, channels and agencies.

Typical timeline: 5–9 weeks

Product & packshot model

A model tuned on your actual products so generated scenes show the real thing: correct proportions, materials, labels and finish.

What’s included

  • Controlled product capture or asset audit
  • Likeness tuning and validation
  • Scoring for product accuracy, not just style
  • Scene and packshot prompt set
  • Likeness
  • Packshots
  • Accuracy

Best forRetail and consumer brands with large catalogues and short seasons.

Typical timeline: 4–8 weeks

Style adapter library

A set of small, swappable adapters for sub-brands, campaigns or markets, so one base model serves several looks without retraining everything.

What’s included

  • Adapter per sub-brand, campaign or market
  • Composition and conflict rules between adapters
  • Scoring per adapter
  • Naming, versioning and documentation
  • Adapters
  • Sub-brands
  • Swappable

Best forHouses of brands and groups running several identities from one team.

03Evaluation & guardrails

3 services
Typical timeline: 3–5 weeks

Brand fidelity evaluation

A fixed scoring set for palette, mark use, typography, composition and tone, so releasing a model version stops being a matter of opinion.

What’s included

  • Scoring criteria written with your brand team
  • Held-out evaluation set
  • Automated scoring calibrated to human review
  • A score report per model version
  • Evals
  • Regression
  • Per version

Best forBrands already generating at volume with no consistent way to judge it.

Typical timeline: 2–5 weeks

Guardrails & misuse testing

Limits on what the tools may produce, and adversarial testing for prompt injection, competitor marks, named people and content the brand must never make.

What’s included

  • Input and output guardrails
  • Prompt-injection and jailbreak testing (OWASP LLM01)
  • Blocklists for marks, people and claims
  • Findings, fixes and a retest
  • Guardrails
  • Red team
  • OWASP LLM

Best forAny brand putting a generative tool in front of a wider team.

Typical timeline: 3–6 weeks

Human approval & output log

An approval step before generated work is used, with the prompt, model version, inputs and approver recorded against every output.

What’s included

  • Approval queue with named owners
  • Model version and settings recorded per output
  • Searchable output log
  • Export for audit or a rights query
  • Human in the loop
  • Audit trail
  • Slack

Best forRegulated brands and teams where legal signs off on creative.

04Serving & care

4 services
Typical timeline: 4–8 weeks

Model serving & registry

Your tuned models deployed behind an API, with versions tracked, cost and latency limited, and a rollback to the previous version when needed.

What’s included

  • Serving endpoint with authentication
  • Model registry and versioned releases
  • Cost, rate and latency limits
  • Rollback and incident runbook
  • API
  • Versioned
  • Cost limits

Best forTeams with a tuned model only one specialist can currently run.

Typical timeline: 4–8 weeks

Integration into your design tools

The models reachable where the work happens: a plugin in your design tool, an action in your content system, or a call from your own pipeline.

What’s included

  • Plugin or app in your design tool
  • API for your content and campaign systems
  • Authentication and per-team permissions
  • Usage documentation for each team
  • Figma
  • CMS
  • API

Best forTeams who will not adopt a tool that lives in a separate tab.

Typical timeline: Monthly, ongoing

Model care & re-tuning

Ongoing care of the brand models: re-tuning when the brand moves, re-scoring each release, and watching cost and drift month to month.

What’s included

  • Scheduled re-tuning and re-scoring
  • Dataset kept current with the brand
  • Cost, latency and drift monitoring
  • Monthly report and backlog
  • Retainer
  • Re-tune
  • Monitored

Best forBrands that now depend on a tuned model for day-to-day delivery.

Typical timeline: 2–4 weeks

Handover & internal ownership

Everything transferred so your team can run and retrain the models: weights, data, scripts, scores and the documentation to do it again.

What’s included

  • Weights, datasets and scripts in your accounts
  • Model cards and training documentation
  • Runbook for re-tuning and release
  • Working sessions with your team
  • Handover
  • Documented
  • Ownership

Best forOrganisations bringing model work in-house after a first build.

Your brief

Tick “Add to brief” on any service, choose a package, then continue. Or enquire about one service directly.

Start a project

Ways to engage. Same team, same standard.

Three ways in, from a two-minute question to a formal RFQ. Each is read in full by the lead for the work, and anything already in your brief goes with it.

Or book a thirty-minute call

What are you sending?

  1. 01

    About 2 minutes4 required answers

    For a first conversation, a press request, or anything that does not need a scope yet.

    You get A reply from a lead, not a sales queue

  2. 02Most useful

    About 8 minutes5 short steps

    Goals, audiences, a budget band and timing. Enough for us to come back with a shape, not only questions.

    You get Options and a first scope after one call

  3. 03

    About 15 minutesYour documents attached

    Your pack, your deadlines, and the procurement and security rules the work must meet.

    You get Receipt confirmed and a named bid lead

How it is priced

Each package shows how it is priced. Every engagement starts with a written scope and a quote agreed before work begins.

  • Typical length
    1–3 weeks
    Pricing
    Fixed fee

  • Typical length
    4–12 weeks
    Pricing
    Fixed price

  • Typical length
    3–9 months
    Pricing
    Fixed price per milestone

  • Typical length
    Ongoing · 6-month minimum
    Pricing
    Monthly fee
Compare what each package includes
What every engagement package includes, and who it suits
PackageEvery engagement includesBest for
SprintOne fixed question, answered in one to three weeks.
  • Scope and outcome agreed before day one
  • One senior lead and the specialists the question needs
  • A working review every week
  • A decision-ready output, not a status deck
 Discovery, a diagnostic, a prototype or a decision you need to make soon
ProjectA defined scope, delivered for a fixed price.
  • Statement of work with deliverables and acceptance criteria
  • A named project lead and a fixed team
  • A shared plan with dated checkpoints
  • Source files and IP transferred on delivery
 Work you can describe up front: an identity, a system, a set of tools
MilestoneA larger build, split into gated phases you approve and pay for one at a time.
  • Phases with their own scope, output and sign-off
  • A go or no-go review at every gate
  • Re-planning between phases as you learn
  • Payment tied to accepted milestones
 Programmes too big to fix in one contract, and teams that want control at each step
RetainerReserved monthly capacity to run, improve and extend what we built.
  • A reserved block of team time every month
  • Agreed response times for requests and fixes
  • A monthly review and a rolling backlog
  • Continuous improvement, not just upkeep
 Brands and products after launch that need a steady team without hiring one

Questions

Asked plainly, answered plainly.

What buyers ask before Brand AI Tools work. Anything else, ask the team directly.

Ask the team
How is this different from Brand AI Tools in Brand Design?

Brand Design builds the tooling around a brand system: the brand check, the template engine, the asset pipeline. AI Design builds the model layer those tools call: the dataset, the tuning, the fidelity scoring, the serving and the re-tuning. Many clients buy both, and the difference is which side leads.

How much material do we need to tune a model?

Less than most people expect for style, more than most expect for likeness. A style adapter can work from a few dozen consistent, well-labelled images; product or person likeness needs controlled, varied captures. We test on a small set first and tell you plainly if the material is not there.

Who owns the trained model?

You do. Weights, adapters, datasets, prompts and logs are delivered into your accounts. We do not retain them and we do not train anything else on them.

What happens when the brand changes?

The dataset and the fidelity set are versioned with the brand. A refresh re-tunes on the new material and re-scores against both the old and the new rules; the previous version stays available until the new one passes.

Can the model generate a real person or a competitor’s brand?

No. Guardrails block named people without recorded consent, and third-party marks, and every output is logged. Misuse testing is part of release rather than an afterthought.

Also in AI Design

A tuned model earns its keep inside a studio and a plan.

Each shares the models, evaluation sets and approval rules built here.

All of AI Design

02 / 04 · Volume without losing the eye

AI Content Studio

AI creatives, films and edits — model selection, automation and creative direction, with tools like Runway, Veo and ElevenLabs.

Explore AI Content Studio

01 / 04 · Interfaces for systems that guess

AI Application Design

Assistants and multimodal, agentic experiences built across Gemini, OpenAI, Anthropic, and beyond.

Explore AI Application Design

04 / 04 · Adoption that survives the pilot

AI Strategy & Consulting

Bringing AI into the brand and marketing ecosystem through pilots and adoption roadmaps built to stick.

Explore AI Strategy & Consulting

Let’s build what happens next.

Tell us what you’re building. We’ll answer straight.

Book a discovery call

Three ways to start

Every engagement starts with a written scope and a quote agreed before work begins.

Choose one of the three ways above
See the work Explore our services Industries we serve Meet the team Email us Read our approach